X-ray photon-counting data correction through deep learning
A method for x-ray photon-counting data correction. The method includes generating, by a training data generation module, training input spectral projection data based, at least in part, on a reference spectral projection data. The training input spectral projection data includes at least one of a pulse pileup distortion, a charge splitting distortion, and/or noise. The method further includes training, by a training module, a data correction artificial neural network (ANN) based, at least in part, on training data. The data correction ANN includes a pulse pileup correction ANN, and a charge splitting correction ANN. The training data includes the training input spectral projection data and the reference spectral projection data.
1. A method for x-ray photon-counting data correction, the method comprising:
generating, by a training data generation module, training input spectral projection data based, at least in part, on a reference spectral projection data, the training input spectral projection data comprising at least one of a pulse pileup distortion, a charge splitting distortion, and/or noise; and
training, by a training module, a data correction artificial neural network (ANN) based, at least in part, on training data, the data correction ANN comprising a pulse pileup correction ANN, and a charge splitting correction ANN, the training data comprising the training input spectral projection data and the reference spectral projection data.
2. The method of claim 1 , wherein the training is performed in a Wasserstein generative adversarial network (WGAN) framework.
3. The method of claim 1 , further comprising:
generating, by the pulse pileup correction ANN, an intermediate estimate based, at least in part, on the training input spectral projection data; and
determining, by a guidance loss circuitry, a guidance loss based, at least in part, on the intermediate estimate, and based, at least in part, on a charge splitting distorted target, the charge splitting distorted target included in the training data.
4. The method of claim 3 , further comprising:
generating, by the charge splitting correction ANN, a pulse pileup and charge splitting corrected estimate based, at least in part, on the intermediate estimate; and
determining, by a correction loss ANN, a correction loss based, at least in part, on the pulse pileup and charge splitting corrected estimate, and based, at least in part, on a ground truth target, the ground truth target included in the training data.
5. The method of claim 3 , further comprising:
generating, by the charge splitting correction ANN, a pulse pileup and charge splitting corrected estimate based, at least in part, on the intermediate estimate; and
determining, by a discriminator ANN, a WGAN loss based, at least in part, on the pulse pileup and charge splitting corrected estimate, and based, at least in part, on a ground truth target.
6. The method of claim 1 , further comprising correcting, by the data correction ANN, an actual input spectral projection data to reduce or eliminate one or more of a pulse pileup effect, and/or a charge splitting effect.
7. The method of claim 1 , wherein the training comprises determining a loss function, the loss function comprising a guidance loss, a correction loss, and a discriminator loss.
8. A data correction artificial neural network (ANN) for x-ray photon-counting data correction, the data correction ANN comprising:
a pulse pileup correction ANN; and
a charge splitting correction ANN,
the pulse pileup correction ANN and the charge splitting correction ANN trained based, at least in part, on training data, the training data comprising training input spectral projection data and a reference spectral projection data, the training input spectral projection data generated based, at least in part, on the reference spectral projection data, the training input spectral projection data comprising at least one of a pulse pileup distortion, a charge splitting distortion, and/or noise.
9. The data correction ANN of claim 8 , wherein the training is performed in a Wasserstein generative adversarial network (WGAN) framework.
10. The data correction ANN of claim 8 , wherein the pulse pileup correction ANN is configured to receive actual input spectral projection data and to provide an intermediate estimate, the intermediate estimate corrected for pulse pileup effects.
11. The data correction ANN of claim 10 , wherein the charge splitting correction ANN is configured to receive the intermediate estimate and to provide a corrected spectral projection data output, the corrected spectral projection data output corrected for pulse pileup effects and charge splitting effects.
12. The data correction ANN of claim 8 , wherein the reference spectral projection data is generated based, at least in part, on a phantom.
13. The data correction ANN of claim 8 , wherein the pulse pileup correction ANN corresponds to a representation network, and the charge splitting correction ANN corresponds to a convolutional neural network.
14. A deep learning system for x-ray photon-counting data correction, the deep learning system comprising:
a data correction artificial neural network (ANN) comprising a pulse pileup correction ANN, and a charge splitting correction ANN;
a training data generation module configured to generate training input spectral projection data based, at least in part, on a reference spectral projection data, the training input spectral projection data comprising at least one of a pulse pileup distortion, a charge splitting distortion, and/or noise; and
a training module configured to train the data correction ANN based, at least in part, on the training data.
15. The deep learning system of claim 14 , wherein the training is performed in a Wasserstein generative adversarial network (WGAN) framework.
16. The deep learning system of claim 14 , wherein:
the pulse pileup correction ANN is configured to generate an intermediate estimate based, at least in part, on the training input spectral projection data; and
a guidance loss circuitry is configured to determine a guidance loss based, at least in part, on the intermediate estimate, and based, at least in part, on a charge splitting distorted target, the charge splitting distorted target included in the training data.
17. The deep learning system of claim 16 , wherein:
the charge splitting correction ANN is configured to generate a pulse pileup and charge splitting corrected estimate based, at least in part, on the intermediate estimate; and
a correction loss circuitry is configured to determine a correction loss based, at least in part, on the pulse pileup and charge splitting corrected estimate, and based, at least in part, on a ground truth target, the ground truth target included in the training data.
18. The deep learning system of claim 16 , wherein:
the charge splitting correction ANN is configured to generate a pulse pileup and charge splitting corrected estimate based, at least in part, on the intermediate estimate; and
a discriminator ANN is configured to determine a WGAN loss based, at least in part, on the pulse pileup and charge splitting corrected estimate, and based, at least in part, on a ground truth target.
19. The deep learning system of claim 14 , wherein the data correction ANN is configured to correct an actual input spectral projection data to reduce or eliminate one or more of a pulse pileup effect, and/or a charge splitting effect.
20. The deep learning system of claim 14 , wherein the pulse pileup correction ANN corresponds to a representation neural network, and the charge splitting correction ANN corresponds to a convolutional neural network.